The US may be moving toward a standing referee for frontier AI. This week, Bloomberg reported that the government is weighing an independent regulatory agency to vet the most powerful AI models — one modeled on FINRA, the securities industry's self-regulator, and reporting to the SEC.
Who is behind it
The proposal was developed with help from Treasury Secretary Scott Bessent, and is now in the hands of White House Chief of Staff Susie Wiles, who is reviewing it. It has not yet been presented to President Trump — an important marker that this is an internal proposal, not policy.
How it would work
Under the plan, frontier labs would voluntarily submit models for review roughly 30 days before public release. The screening would target specific dangerous capabilities: deception, the ability to help create bioweapons, and malicious hacking or cyber skills. A FINRA-style body would replace today's improvised, case-by-case interventions with a repeatable process.
Why now
Recent months made the ad-hoc approach visible. Anthropic saw its Fable 5 and Mythos 5 models temporarily disabled from user access under US export controls before they were lifted around the start of July, and OpenAI limited the rollout of its latest Sol model, part of the GPT-5.6 family, at the government's request. A standing agency would formalize that pre-release gatekeeping rather than improvising it each time.
The caveats
Nothing is settled: no legislation has been drafted, no funding allocated and no timeline announced, and submission would be voluntary. The structure is already drawing criticism — former White House AI czar David Sacks reportedly called an SEC-reporting arrangement "concerning." It should not be confused with a separate, industry-run standards body proposed days earlier by DeepMind's Demis Hassabis; this one is a government-linked regulator, not a lab-funded club.